Access Considerations for Generative AI Systems Beyond Release
Summary
Irene Solaiman urges policymakers to look past “open versus closed” and instead judge AI systems by how costs, technical usability, and utility determine who can actually use these systems
SESSION Transcript
Hey everyone. So my name's Irene. I'm the head of global policy at a place called Hugging Face. It's this emoji, it's that one.
And let's see. So I published this paper two years ago now, back in the old times of 2023, and it was really challenging this idea of open versus closed, that instead there's this spectrum, there's this gradient, and I put these kinds of trade-offs at either end.
One part that I didn't think through back in those old times was where it's bold, it says broader integration at the bottom. Now we've seen particularly with smaller open-weight models like DeepSeek smaller models, that once you have some open-weight systems, they can often actually be broader integrated into our global ecosystem.
So now, sitting from my role, I've really just had more of a sense of not just a race for AGI, not just a race for more powerful systems, but really a race for this ecosystem and where open weights fits in. That being said, we've heard some concerns about open-weight innovation and safety. I'll be hosting a discussion later this evening. If you're around, we'd love to have you there.
But particularly something that we've witnessed on Hugging Face, but generally in the open, open source, open-weight world is that sometimes there are open-weight models, but people don't have the resources to be able to deploy or host them. This is particularly true of these large models. I made this meme myself, please compliment me on it later. I wrote this paper a couple months ago.
It's gone through several rounds of peer review, which is part of why I feel comfortable sharing with you today, of why it's so important to look past open versus closed and more to available versus accessible. This is regardless of whether a model and this is looking through systems. But we'll talk about models since we're talking about open weights right now. If a model is available, this might be through interfaces.
How many of you in this room have used a language model and how many of you use that through a text box, some sort of app interface? How many of you did it hosting it yourself using Python? Far fewer. So sometimes these interfaces can make it really more accessible, which is part of why we see such a boom in AI technology.
So one example that I would give for accessible versus available is this is not to pick particularly on ChatGPT. It's just an incredibly accessible system. It enables so many people without technical backgrounds to use that system. On January 1st, one of the attackers who blew up the Cybertruck in front of the Trump Tower workshopped his attack with ChatGPT.
Again, this is not because of the model. It's really easier for somebody who's non-technical to be able to work with that model. On the flip side, for example Llama3 405B, sometimes that's actually too computationally expensive for people to actually meaningfully work with it. The three areas that I've seen have actually helped people make available systems accessible are resourcing.
So that can be compute or per-token cost. And that means maybe fewer researchers like we've seen with larger open-weight models can use it, but also fewer harmful actors can use it. Technical usability like these interfaces or APIs and then utility. I realize the wording is kind of hard, but something like languages. So if you have a low resource language available in a language model, maybe you open up to that population and that's incredible.
But meaningfully monitoring low resource languages can be really hard. We've seen devastating consequences with social media. And when you make systems generally more accessible, they scale easier. This adds to considerations with distribution, not just distribution of the system, but if it's outputs.
Outputs are only as impactful as who they reach. And it also just makes it really hard to monitor and manage at a really broad scale, regardless of whether that system is open or closed. When you have a system via API that is receiving millions of queries a day, you don't want to the Internet speakers nerf your model and make it really inaccessible to use, unavailable for commercial use cases.
But you're trying to navigate these many different safety considerations. So throughout this paper I encourage you to take a look at it. We compare models Llama, DeepSeek, GPT-4o and Claude, all of which were at the top of Stanford's HELM leaderboard in early 2025 when I did this work. And generally a lot of the considerations are pretty similar outside of per token cost for DeepSeek.
And then we look at for example usability, a lot of them have APIs through themselves or a third party hosting service. We've seen a rise in licenses and customized licenses, but even more permissive. So Qwen family, for example, I believe it was the Qwen 2 family had staged licenses with permissive for the smaller ones and customized for the larger ones.
And then for utility it's actually quite similar. So this is where I want to shift that framing, have you think more about it. That being said, not everything's about accessibility. There's changes over time like DeepSeek, there's economic and concentration of power and scientific considerations.
But I will leave you with this to think deeper about. What does it mean for us to actually have more accessible systems? And what does it mean for the people who benefit and who are harmed? Hope that was helpful.
Happy to discuss it later.
And let's see. So I published this paper two years ago now, back in the old times of 2023, and it was really challenging this idea of open versus closed, that instead there's this spectrum, there's this gradient, and I put these kinds of trade-offs at either end.
One part that I didn't think through back in those old times was where it's bold, it says broader integration at the bottom. Now we've seen particularly with smaller open-weight models like DeepSeek smaller models, that once you have some open-weight systems, they can often actually be broader integrated into our global ecosystem.
So now, sitting from my role, I've really just had more of a sense of not just a race for AGI, not just a race for more powerful systems, but really a race for this ecosystem and where open weights fits in. That being said, we've heard some concerns about open-weight innovation and safety. I'll be hosting a discussion later this evening. If you're around, we'd love to have you there.
But particularly something that we've witnessed on Hugging Face, but generally in the open, open source, open-weight world is that sometimes there are open-weight models, but people don't have the resources to be able to deploy or host them. This is particularly true of these large models. I made this meme myself, please compliment me on it later. I wrote this paper a couple months ago.
It's gone through several rounds of peer review, which is part of why I feel comfortable sharing with you today, of why it's so important to look past open versus closed and more to available versus accessible. This is regardless of whether a model and this is looking through systems. But we'll talk about models since we're talking about open weights right now. If a model is available, this might be through interfaces.
How many of you in this room have used a language model and how many of you use that through a text box, some sort of app interface? How many of you did it hosting it yourself using Python? Far fewer. So sometimes these interfaces can make it really more accessible, which is part of why we see such a boom in AI technology.
So one example that I would give for accessible versus available is this is not to pick particularly on ChatGPT. It's just an incredibly accessible system. It enables so many people without technical backgrounds to use that system. On January 1st, one of the attackers who blew up the Cybertruck in front of the Trump Tower workshopped his attack with ChatGPT.
Again, this is not because of the model. It's really easier for somebody who's non-technical to be able to work with that model. On the flip side, for example Llama3 405B, sometimes that's actually too computationally expensive for people to actually meaningfully work with it. The three areas that I've seen have actually helped people make available systems accessible are resourcing.
So that can be compute or per-token cost. And that means maybe fewer researchers like we've seen with larger open-weight models can use it, but also fewer harmful actors can use it. Technical usability like these interfaces or APIs and then utility. I realize the wording is kind of hard, but something like languages. So if you have a low resource language available in a language model, maybe you open up to that population and that's incredible.
But meaningfully monitoring low resource languages can be really hard. We've seen devastating consequences with social media. And when you make systems generally more accessible, they scale easier. This adds to considerations with distribution, not just distribution of the system, but if it's outputs.
Outputs are only as impactful as who they reach. And it also just makes it really hard to monitor and manage at a really broad scale, regardless of whether that system is open or closed. When you have a system via API that is receiving millions of queries a day, you don't want to the Internet speakers nerf your model and make it really inaccessible to use, unavailable for commercial use cases.
But you're trying to navigate these many different safety considerations. So throughout this paper I encourage you to take a look at it. We compare models Llama, DeepSeek, GPT-4o and Claude, all of which were at the top of Stanford's HELM leaderboard in early 2025 when I did this work. And generally a lot of the considerations are pretty similar outside of per token cost for DeepSeek.
And then we look at for example usability, a lot of them have APIs through themselves or a third party hosting service. We've seen a rise in licenses and customized licenses, but even more permissive. So Qwen family, for example, I believe it was the Qwen 2 family had staged licenses with permissive for the smaller ones and customized for the larger ones.
And then for utility it's actually quite similar. So this is where I want to shift that framing, have you think more about it. That being said, not everything's about accessibility. There's changes over time like DeepSeek, there's economic and concentration of power and scientific considerations.
But I will leave you with this to think deeper about. What does it mean for us to actually have more accessible systems? And what does it mean for the people who benefit and who are harmed? Hope that was helpful.
Happy to discuss it later.
